►
Description
Come hang out with Josh Rosso as he does a bit of hands on hacking of Kubernetes and related topics. Some of this will be Josh talking about the things he knows. Some of this will be Josh exploring something new with the audience. Come join the fun, ask questions, comment, and participate in the live chat!
This week we'll be looking at flagger (https://flagger.app/). This tool enhances application rollout capabilities by provide release validation and configurable routing strategies.
- 00:00:00 - Welcome to TGIK!
- 00:02:05 - Week in Review
- 00:19:00 - Flagger Introduction
- 00:24:32 - Contour Deploy
- 00:28:26 - Flagger Deployment
- 00:53:06 - Canary Test (success and failure)
- 01:16:12 - A/B Test
- 01:29:59 - Summary and Wrap-up
TGIK episode directory (GitHub): https://github.com/vmware-tanzu/tgik/tree/master/episodes/106
A
Hey
everyone
welcome
how's.
It
going
welcome
to
episode
106
glad
to
have
you
all
really
excited
to
be
back,
as
always,
if
you're
just
joining
us
feel
free
to
say
hello
in
the
chat.
If
you
want
to
say
where
you're
tuning
in
from
that's
always
welcome,
see
who
we've
got
joining
us
so
far,
Alex
how's
it
going
welcome
I'll
lead
grab.
Glad
to
have
you
back,
looks
like
we've
got
Duffy
with
us
today,
which
will
be
great,
hey,
Duffy,
feel,
like
I
haven't
seen
you
all
week.
A
You
must
have
been
really
busy
Martin
from
the
Netherlands.
Welcome
Martin
glad
to
have
you
here,
Rodolfo
ready
to
learn
good,
happy
Friday
as
well
Matty
glad
to
have
you
back,
I
think
this.
This
issue
and
recommendation
came
from
you
Matty.
So
thanks
for
that
inside
of
github,
we
really
appreciate
it
so
riche
we
got
you
from
hamburg.
A
Welcome
Tim
in
from
San
Francisco
welcome,
Tim
glad
to
have
you
ABC
welcome
back,
always
happy
to
have
you
given
us
some
good
insight,
Nick,
yep,
Nick
Nick
knows
that
my
my
work
setup
is
in
my
closet,
because
it's
quiet
in
here
and
very
easy
to
focus.
So
thanks
Nick,
welcome
glad
to
have
you,
man,
Grigore
from
Moldova
Moldova
awesome.
Welcome!
Welcome
glad
to
have
you
Steve,
hey
how
you
doing
we're
gonna,
be
they
were
gonna,
be
using
some
contour,
today's
Steve.
A
A
Mizzou
fer
from
Dubai
welcome
glad
to
have
you
all
right.
Everyone
welcome.
So,
as
people
are
still
filtering
in,
let's
take
a
little
bit
of
time
and
go
over
the
news,
what
do
you
say
so
I
see
a
lot
of
familiar
faces,
but
for
those
of
you
who
maybe
aren't
as
familiar
with
TGI
K,
one
of
our
starting
points
is
to
talk
about
what's
going
on
in
kubernetes
this
week,
so
we're
gonna
kind
of
kick
that
off
here,
George,
as
always,
has
filled
us
in
with
some
really
really
good
detail
to
kind
of
go
over.
A
So,
let's
start
off
with
the
weekend
review
and
the
core,
so
1.18
we're
getting
closer
and
closer.
There
are
50
enhancements
that
are
being
tracked.
I
actually
want
to
show
you
all
some
of
the
ones
that
I'm
most
excited
about
and
if
you're
not
aware
of
where
the
enhancements
spreadsheet
is,
you
might
find
that
pretty
interesting.
A
A
We
have
in
our
hack
MD
around
patch
release
updates
as
well,
but
let's
take
a
moment
to
talk
about
some
of
the
the
1.18
things
that
are
going
on
here
so
and
I
will
just
click
on
this
cool,
so
inside
of
our
hack,
MD,
which
looks
like
George
may
have
just
put
the
notes.
Link
in
chat
so
feel
free
to
check
that
out
too.
There
is
a
link
in
here
which
is
the
enhancements
tracking
spreadsheet,
which
is
a
pretty
cool
way
to
see
what
exactly
is
going
on
for
the
1.18
release.
A
It's
got
some
status
information
I
like
this
enhancements
tab,
because
I
can
tell
whether
it's
something
that
is
being
graduated
or
if
it's
something
that's
kind
of
a
net
new
feature
and
all
that
good
stuff
so
looks
like
there's
lots
and
lots
of
cool
support
for
things
coming
in
a
lot
of
new
updates
and
so
on.
So
do
check
this
out
if
you're
curious,
what
118
is
going
to
have
inside
of
it,
I
pulled
some
of
my
favorites
out.
I
figured
I'd
share
them
with
you
all.
A
A
As
long
as
my
cluster
didn't
die
on
me,
you
can
do
cube
cuddle,
get
events
and
it
will
get
events
for
the
namespace
that
you're
currently
looking
at,
which
is
which
is
cool,
works
pretty
well,
but
there's
definitely
some
limitations
and
I
think
you
know,
as
people
have
been
building
more
and
more
extensions
for
kubernetes
and
kind
of
features
around
kubernetes.
You
know
they've
been
relying
on
using
the
events
API
a
little
bit
more,
which
allows
them
to
kind
of
give
these
more
kind
of
discrete
actions
that
happen
rather
than
just
pumping
everything
through
logs.
A
So
from
the
first
look
and
I've
by
no
means
tested
out,
it
looks
like
some
of
their
intentions
are
to
provide
streaming
of
events
through
this
command.
More
fancy
filter
type
capability,
on
top
of
it
being
able
to
watch
particular
objects
based
on
events.
So
it's
looking
pretty
cool
I
think
this
will
be
a
really
exciting
one,
especially
for
those
of
us
who
get
in
kind
of
the
the
debugging
kind
of
workflow
and
events.
A
Often
times
have
the
information
we
need,
but
we're
trying
to
figure
out
where
to
get
to
the
events,
how
to
filter
through
all
the
massive
sets
of
events
that
might
be
happening
in
a
namespace
and
and
so
on.
So
this
is
kind
of
a
cool
one
to
check
out
in
118
and
then
another
one
speaking
of
debug.
Those
of
you
who
spend
a
lot
of
time,
debugging
like
myself
I,
feel
like
that's
all
I
do.
Nowadays
there
is
a
new
debug
capability
being
added
to
cube
cuddle
as
well.
A
It
appears
so
at
first
glance
of
this
one
and
I
actually
didn't
know
if
this
one
until
today,
so
thanks
George
for
posting,
that
it
looks
like
it's
gonna
make
use
of
the
ephemeral
containers,
which
is
kind
of
a
cool
concept.
We're
inside
of
your
pod.
In
theory,
you'd
be
able
to
set
up
an
ephemeral
container
inside
of
that
pod,
which
could
potentially
have
some
like
debugging
tools
and
things
that
would
allow
you
to
better
debug.
What's
going
on
with
your
app
and
so
what
this
could
help
us
with.
A
So
you
know
when
we
have
these
really
minimal
pods
running
the
ability
to
leverage
ephemeral
containers
to
attach
some
type
of
debugging
tool
set
in
that
pod,
which,
as
you
know,
containers
in
a
pod
share
namespaces
rights
process,
namespaces,
Network,
namespaces,
all
that
good
stuff.
So
you
know,
like
I,
said:
I
haven't
dug
too
deep
into
this,
but
kind
of
as
a
description
description
implies.
You
know
a
lot
of
times
we're
doing
like
exact
import
forwards
and
all
kinds
of
stuff.
A
Maybe
this
could
give
us
a
really
clean
way
to
kind
of
get
inside
how
that
pods
thinking
and
see
what's
going
on
at
the
at
the
container
level,
so
pretty
cool
pretty
excited
about
it,
and
then
one
more
I
was
gonna
share
with
you,
which
I
thought
was
pretty
interesting
as
well.
I
need
to
keep
up
with.
What's
going
on
in
kubernetes,
I
feel
like
I,
just
wait
till
the
release
comes
out
and
then
get
up
to
speed,
but
all
this
is
so
interesting.
A
A
There's
no
true
notion
of
sidecar
right.
So
let
me
give
you
a
more
concrete
example:
let's
say,
you're
running
a
service
match
right
and
your
sidecar
in
the
service
match
for
most
of
your
workloads
is
likely
to
be
something
like
on
volume
right.
So
by
having
this
notion
of
setting
up
a
sidecar
container,
it's
gonna
actually
change
how
some
of
the
life
cycle
pieces
work.
A
So
when
you
kind
of
shut
down
and
start
up
so
trying
to
see
if
I
could
find
the
description
here
where
it
talks
about
the
the
startup
process,
let's
see
it
had
a
pretty
good
description
of
it:
okay,
yeah!
Actually
this
was
it
right
here.
So
an
it
containers
start
into
containers
finish
and
then
our
sidecar
start
right
side
cars
become
ready
and
then
the
containers
themselves
start.
A
I've
done
some
really
gross
things
frankly
in
kubernetes
to
make
sure
things
kind
of
paired
up
well
and
I
think
this
will
give
us
a
little
bit
of
a
cleaner
conception
of
like
this
is
my
primary
at
pod
and
here's
some
of
the
sidecar
pods,
such
as
the
ones
that
are
sorry
I
shouldn't
be
saying
pods,
but
sidecar
containers
such
as
logging,
proxying
and
so
on.
That's
gonna
make
my
workload
function
well
in
this
environment,
so
pretty
cool
yeah,
so
I
just
wanted
to
share
those
three
with
you.
A
The
introduction
of
sidecar
containers,
the
cube
cuddle,
debug
command
and
the
cube
cuddle
events
command.
So
for
our
show
notes,
if
you
haven't
checked
those
out,
get
a
full
list
of
all
that's
going
on
in
118,
so
you're
ready
when
it
comes
out
with
some
of
the
new
features
cool
all
right
so
who
do
we
got
Alex
so
alex
is
on
my
team
at
VMware
Alex
great
to
see
you,
man
welcome.
Welcome
timmi,
also
on
my
in
my
group
as
well
yeah
you're
right
to
me.
Where
is
my
hat?
That's
a
great
question!
A
Unfortunately,
I
don't
know,
but
you're
right
I
do
feel
a
little
a
little
weird
without
it
now,
especially
now
that
you
pointed
it
out
thanks
to
him
so
cool
cool,
all
right
and
while
lead
keep
up
with
the
cloud
native
pod.
Let's
last
episode,
yep
very
cool
I'm
glad
you've
checked
out
the
pod.
Let's
podcast
wall-e,
that's
a
new
podcast
we're
trying
to
put
out
weekly
about
cloud
native
stuff.
Stefan!
Welcome!
Welcome
glad
to
have
you
all
right,
so
let's
get
a
little
bit
further
and
then
we'll
we'll
get
into
some
goodness
around
flagger.
A
A
We
have
an
app
we're,
gonna
spin
up
a
kubernetes
cluster
for
that
app,
or
maybe
you
know,
a
couple
of
micro
services
that
make
up
the
app
and
it's
going
to
run
on
this
cluster
and
then
on
the
other
side
of
this
bactrim.
We
have
the
large
shared
clusters
which
I
think
when
kubernetes
first
became
popular.
At
least
this
was
my
experience
is
that
everyone
was
shooting
for
this
large
shared
cluster
model
right
and
you
know
I
think
I,
think
a
lot
of
that
and
I
don't
really
know
the
inner
workings
of
Borg.
A
My
experience
and
I
think
this
does
a
pretty
good
job
of
kind
of
talking
about.
It
is
a
lot
of
folks
who
started
there
realize
that
that's
really
hard
to
do,
and
at
minimum
they
split
up
their
different
environments
into
different
clusters
right
so
dev
prod
test
QA,
whatever
million
you
know
qualifiers,
they
have
for
their
different
environments.
They'd
split
those
up
at
least
to
get
some
segregation
there
and
then,
as
time
went
on.
A
I'd
argue
that
even
more
and
more
kind
of
splitting
up
of
clusters
happen
with
a
lot
of
the
people
that
I
worked
with
where
they
were
running
more
smaller
clusters
and
I
think
it
was
kind
of
twofold
one.
It
was
just
experience
with
kubernetes
and
how
it
can
be
hard
to
run
one
massive
large
cluster,
but
to
also
like
lifecycle.
Management
of
kubernetes
is
getting
easier
over
time.
So
as
we're
talking
more
about
Federation
and
service
routing
capabilities,
you
probably
all
are
very
aware
of
cluster
API
when
we
actualize
that
into
something
that's
really
stable.
A
You
know
the
beauty
of
just
making
an
API
request
and
clusters
showing
up
that
can
really
enable
us
to
have
you
know
more
multi
cluster
facilities
without
as
much
overhead
of
managing
each
individual
cluster
but
anyways.
That's
that's.
My
soapbox
I
would
check
this
article
out.
I
read
through
it
very
very
quickly
and
I
think
they
bring
up
some
really
great
points.
You
know
they
go
through
like
one
large
shared
cluster
and
like
everything,
it's
just
a
list
of
trade
offs
right.
It's
what
trade
offs
are
right
for
you.
A
A
So
when
you
have
like
a
multi-tenant
cluster,
like
imagine
that
you're
a
company,
that's
hosting
customers
on
a
shared
cluster,
you
know
the
tenancy
constraints
that
you
need
to
put
in
place
and
the
isolation
you
need
to
put
in
place
are
pretty
serious
in
that
case,
they're
a
lot
more
serious
and
they'd,
be,
if,
like
you
know,
Duffy
and
I
are
just
running
our
app
side-by-side
and
some
cluster
into
some
degree.
We
we
sort
of
trust
each
other
right.
A
You
know
it
can
be
really
tough
to
to
have
a
security
isolation
and
these
these
larger
clusters.
Then
they
go
over
the
trade
offs
for
kind
of
the
single
use
small
clusters.
This
would
be
almost
like
a
cluster
per
app
and
then
arrests,
sorry,
that
was
in
section
3
here,
cluster
per
app
and
then
they've
got
the
cluster
per
environment
model.
So
you
know
like
everything.
This
is
a.
This
is
a
spectrum
where
you
choose
where
to
land
on
it,
based
on
your
trade-offs
in
your
environment,
but
nonetheless,
it's
really
detailed
check
it
out.
A
They
did
an
awesome
job,
so
great
work,
Daniel.
If
you
ever
see
this,
it's
a
really
good,
write-up
cool.
So
another
really
great
one
I
feel
like
this
has
surfaced
a
couple
times
now:
CPU
limits
and
aggressive,
throttling
and
kubernetes.
Excuse
me
phii's.
Hopefully,
I'm
saying
it
somewhat
correctly
talks
a
little
bit
about
some
of
the
underpinnings
of
how
resource
allocations
work
in
kubernetes.
A
So
not
only
might
this
give
you
a
consideration
depending
on
how
you've
got
your
cluster
set
up
for
CPU
throttling,
but
it
talks
about
some
of
the
kind
of
in
the
weeds
things
that
we
don't
always
know
about.
You
know
one
of
the
things
in
kubernetes
that
I
think
a
lot
of
us
are
aware
of.
Is
you
have
the
ability
to
request
a
certain
amount
of
CPU
and
you
have
the
ability
to
limit
a
certain
amount
of
CPU?
And
you
know
sometimes
it's
simplified
and
we
say
like
okay.
A
Well,
requests
are
for,
like
the
kubernetes
scheduler,
to
figure
out
where
to
put
something
and
how
much
it
should
conceptually
carve
out
and
then
the
limit
is
like
enforced
at
the
runtime
level.
This
will
actually
like
be
throttled
at
a
certain
level
and
if
your
CPU
hits
that
will
throttle
it
down
now
in
actuality,
it's
it's
a
little
bit
different
than
that
at
runtime,
at
least
for
CPU.
Both
of
these
things
do
get
respected.
A
You
know,
as
an
example,
let's
see
if
they've
got
them
here.
Cpu
shares
is
what
I'm
looking
for
yeah.
So
the
CPU.
Let's
see
if
you
mentioned
the
ZFS
anywhere
yeah
there
we
go
so
inside
of
requests,
it's
going
to
go
in,
and
it's
going
to
provide
kind
of
like
this.
Certain
amount
of
shares
you
get
inside
of
the
inside
of
the
runtime
right
and
then
with
limits.
A
It's
going
to
use
this
thing
called
a
completely
fair
scheduler
to
actually
throttle
you,
so
the
notion
being
like
if
you're,
if
your
node
comes
under
contention,
you
could
have
a
higher
limit
than
what
you
requested
and
if
that
CPU
is
available,
it's
actually
gonna.
Let
the
CPU
go
higher,
at
least
up
to
that
limit
bound
right.
A
But
if
the
host
comes
under
contention,
it'll
start
kind
of
throttling
down
the
CPU,
and
what
this
article
goes
over
is
basically
the
implications
of
that
the
calculation
of
the
the
CPU
limits
themselves
and
how
kind
of
those
get
calculated.
So
it
talks
about,
unlike
CPU
shares.
This
quota
is
based
on
a
time
period
and
how
that
time
period
calculation
happens
so
check
this
out.
It's
super
interesting.
A
It
shows
how
under
resource
contention,
the
error
rates
can
actually
go
up
in
the
presence
of
CPU
limits,
so
in
this
article
and
I
believe
they
call
it
out
directly,
but
I
believe
the
too
long
didn't
read.
Is
they
modified
whether
the
cubelet
was
even
respecting
CPU
limits
at
all?
So
it's
really
really
cool.
There's
another
article,
that's
very
similar
to
this
one.
If
you
want
to
dive
deeper
into
it,
it
was
done
by
Zhu,
Lando
I'll.
Look
it
up
after
this
video
and
put
it
in
the
show
notes.
A
They
actually
had
a
presentation
where
they
talked
a
little
bit
about
the
implications
they
found
from
running
CPU
limits
and
how
it
impacts
the
completely
fair,
scheduler
and
and
all
that
good
stuff.
So
really
cool
check
it
out
pretty
pretty
in
the
weeds,
but
good
information
to
know
cool
and
then
one
more
that
I'll
just
mention
and
then
we'll
get
right
into
it
is
on
the
is
on
deploying
spark
to
kubernetes.
A
So
a
guide
was
released
from
developer
SH
on
putting
spark
on
cube,
which
a
lot
of
folks
want
to
do,
but
like
some
stateful
Java
apps,
it
can
be
a
little
bit
challenging
to
know
how
to
set
it
up
correctly
or
I.
Guess
just
Java
apps
in
general,
sometimes
they're
a
little
bit
more
complicated
than
those
of
us
that
live
in
a
you
know:
static
binary
world,
so
check
this
article
out.
If
you're
looking
to
run
spark,
they
have
some
cool,
gotchas
and
instructions
on
how
to
set
it
up
and
run
it.
A
So
pretty
cool
stuff,
cool,
all
right.
All
right,
I'm
gonna
check
chat
one
more
time
and
then
we'll
get
right
into
it.
Maddy
anyone
happened
to
know
if
pod
startup
shut
down
or
during
ever
got
implemented,
I'm,
not
sure
offhand
Maggie,
it's
a
good
question.
I,
do
not
that
I
have
not
that
I
have
seen
at
least,
but
it
might
be
a
feature.
A
That's
out
there
and
I'm
just
not
aware
cool
all
right
all
so
we're
gonna
talk
a
little
bit
about
flagger
now,
so
those
of
you
who
are
in
chat
I
know
I
talked
to
some
of
the
folks
in
the
weave
work
slack.
Do
we
have
any
flagger
maintainer
Zoar
we've
work
employees
in
the
chat
who
are
with
us
today,
if
so
feel
free
to
say,
hey
all
right.
Let
me
just
pull
over
my
flagger
there.
We
go
cool.
A
A
Real,
quick,
flagger,
awesome,
Stephane,
oh
yeah
I
should
have
recognized
your
name
thanks
for
joining
us
Devon,
so
those
in
the
chat
Stefan's
here
who's,
a
flagger
maintainer,
so
he'll
be
helping
us
out
and
answering
questions
as
they
as
they
come
up.
Thanks
for
joining
us
today,
all
right
just
getting
my
brain
together
as
we
switch
from
the
news
to
to
flagger
itself.
So
this
is
probably
something
that
most
of
you
who
have
run
flagger
in
the
past
are
sorry.
A
So
those
of
you
who
have
run
kubernetes
in
the
past
probably
know
that
one
of
the
kind
of
complex
things
that
we
want
to
do
is
have
some
more
intricate
rollout
strategies.
You
know
thinking
about
how
we
can
slowly
bring
a
new
service
or
an
updated
service
even
into
production
and
there's
a
couple
things
that
we
oftentimes
consider
here.
A
You
know,
like
the
the
very
kind
of
naive
way
that
we've
tried
to
do
it
in
kubernetes
before,
as
many
of
us
know,
is
we'd
start
off
with
the
idea
of
a
replica
set
right,
which
is
probably
backed
by
a
deployment
and
then
we'd
have
a
service
overtop
of
that
you
know,
kubernetes
101,
stuff
traffic
comes
in
hits
through
the
service
hits
on
the
pod
and
then
oftentimes.
What
we
do
is
we'd
introduced
a
version
to
of
that
app
behind
the
same
service
right
and
then,
since,
at
least
by
default.
A
Those
of
you
who
are
just
using
iptables
to
do
you
know
cube
proxy
based
service
routing.
You
probably
know
that
you
know,
in
the
absence
of
an
ingress
controller
or
anything,
what's
typically
gonna
happen
is
you're
gonna
round-robin
to
those
pods
and
slowly
over
time
you
might
increase
the
pod
count
on
version
2
and
then,
if
things
go
well,
you
decrease
the
pod
count
on
version
a
now
that
very
simplified
model
is
great.
However,
there's,
obviously
a
lot
more
complexity
that
comes
with
doing
this
in
the
real
world.
A
Right
because
you
know
reality
is
we
want
to
slowly
roll
traffic
based
on
maybe
some
type
of
percentage
right.
We
may
want
to
do
some
type
of
validation
logic
to
see
if
things
are
working
correctly,
and
maybe
if
we
can
get
this
fancy
even
rollback,
should
things
blow
up
on
us
and
fail
right.
So
flagger
has
a
bunch
of
different
options
for
doing
progressive
delivery.
Three
of
the
most
common
which
we
run
into
all
the
time
in
the
field
at
least,
are
canary,
which
usually
is
kind
of
introducing
a
new
workload.
A
Slowly
bringing
traffic
over
to
it.
Hopefully
doing
some
type
of
validation
right
canary
in
the
coalmine
is
kind
of
like
where
that
notion
comes
from.
We've
got
a
be
testing,
which
can
be
especially
helpful
for
like
UI
designers
and
people
who
want
to
you
know
say,
introduce
a
certain
version
of
something
to
a
certain
geo
so
that
they
can
see
what
the
results
are
over
time
and
determine
whether
that's
appropriate
before
they
switch
over
and
then
blue
green,
which
is
a
little
bit
more
like
kind
of
a
cutoff
where
we
have.
A
You
know
green,
where
we
have
green
up
with
the
with
the
new
version
and
we're
effectively
able
to
shift
traffic
over
to
it.
But
if
something
goes
wrong,
the
kind
of
roll
back
the
old
version
still
around
is
kind
of
a
I.
Don't
know
a
warm
version
of
the
app
if
you
will
not
completely
killed
so
what's
interesting
about
flagger
is
it
can
fit
into?
A
Based
on
that
change
and
does
the
things
that
are
needed
to
to
make
it
run
well
right
now,
that's
a
super
simplified
naive
explanation
for
git
ops,
but
part
of
that
equation
of
bringing
all
these
things
together
is
to
use
something
like
flagger
to
actually
facilitate
the
rollout
aspect
of
the
app.
So
we
did
some
past
TGI
kaise
as
well.
A
If
you
haven't
seen
them,
I
think
we've
done
I
think
we've
done
flux
and
Argo
right,
I,
someone
cracked
me
in
chat
if
I'm
wrong
there,
but
I
think
we've
done
TGI
Kazon
on
both
of
those,
and
we
are
today
just
gonna
focus
on
the
flagger
aspect
right,
which
is
this
rollout
piece.
So
let's
get
right
into
it.
A
You
want
to
do
this
thing
with
so
maybe
you're
using
sto,
and
you
want
to
kind
of
automate
some
of
the
validation
and
rollout
strategies.
Flagger
can
plug
into
that
link
or
d
nginx
Glu.
So
on
and
today
you
know
selfishly
because
I
know
contour
pretty
well
so
I'm
not
gonna.
Try
to
boil
the
ocean
here
with
deploying
STL
on
top
of
all
of
that,
but
we're
gonna
use
contour
as
an
example.
A
So
flagger
is
gonna
get
deployed
and
with
contour
running,
ideally
we're
gonna
be
able
to
shift
traffic
into
different
services
or
different
deployments
really
on
the
basis
of
what
flagger
has
this
kind
of
roll
out.
So
first
things.
First,
everyone.
Let's
go
ahead
and
deploy
contour
I
have
a
feeling
we
won't
get
super
far
without
contour,
so
inside
of
manifests,
I've
got
a
contour
yeah
Mille
that
I'm
going
to
show
you
all
and
inside
of
this
contour
yeah
Mille.
A
Let's
make
the
text,
maybe
just
a
tiny
bit
smaller,
for
you
all
might
make
it
a
little
bit
readable.
Okay,
so
contour
is
nothing
special
I.
Actually
looked
at
the
flagger
docks
a
little
bit
earlier
and
checked
out
the
contour
version
that
they
were
using.
The
only
difference
is
with
contour.
Maybe
I
should
actually
just
give
a
quick
intro
to
contour.
So
contour
is
an
ingress
controller.
A
It's
one
that
we've
been
working
on
since
a
lot
of
my
group
was
at
hep
do
and
now
at
VMware
it's
been
taken
on
it's
a
it's
a
completely
open
source
project,
CNC
F.
We
have
a
lot
of
different
contributors
on
contour
and
the
idea
with
contour
is
we
use
envoy
as
the
proxy,
okay
and
envoy,
or
reverse
proxy,
no
load-balanced
or
more
or
less,
and
we
have
a
controller
called
contour
that,
based
on
ingress
resources
and
special
resources
that
we
call
HTTP
proxy
I
believe
is
now
what
the
API
is
called.
A
A
This
is
the
Envoy
part
of
the
puzzle
and
how
I've
got
envoy
set
up
is
I'm,
just
gonna
link
it
to
the
host
ports
on
my
machines,
which
I'll
show
you
in
a
moment
so
I'm
not
going
to
use
like
load
balancers
or
anything
like
that
on
top
of
it
we're
just
gonna
route
directly
into
the
machines
in
this
use
case.
So
let's
go
ahead
and
get
contour
deployed,
we'll
open
up
another
buffer
here,
real
quick
and
do
a
cube
cuddle
get
pods
for
the
namespace
project.
Contour
looks
good.
Alright,
so
manifests
contour
dot.
A
A
Alright,
just
check
in
chat
while
it's
starting
out
okay,
cool
flus
and
our
go
cool
thanks
for
posting
those
those
episodes.
Walid
sounds
good
cool.
Oh,
it
has
Argo
changed
a
lot
since
the
last
tea
gik
I
haven't
been
tracking.
It
super
closely
ABC.
We
should.
We
probably
do
a
second
look
kind
of
like
what
we
did
with
psyllium.
If
that
would
interest
you
all
kind
of
another,
look
at
our
go
now
that
stuff
has
changed.
I!
Guess,
honestly,
we
could
probably
take
our
whole
backlog
and
do
a
second
look
at
this
point.
A
Just
more
context,
there
ABC
makes
sense,
Steve
yep
so
to
Steve's
clarification.
So
the
idea
here
is
going
to
be
that
we're
gonna
be
running
the
we're
gonna
be
running
contouring
using
a
Ciardi
called
HTTP
proxy,
which
gives
us
a
bunch
of
capabilities
beyond
just
what
the
default
ingress
api
is,
but
it
is
also
going
to
support
or
also
does
support,
just
ingress
resources.
So
contour
works
natively
with
the
with
the
core
kubernetes
api
as
well.
Should
you
choose
to
use
that
stuff
on
what
version
am
I
using?
A
Let's
just
double
check,
it
should
be
pretty
darn
new.
So
give
me
a
shout
out
if
I'm,
if
I'm,
actively
shooting
myself
in
the
foot
here,
let's
check
out
the
image
version
image
I'm
using
latest,
which
is
probably
1.2,
then
so,
give
me
a
shout:
if
that's
if
the
new
version
would
have
any
issues,
it's
definitely
one
dot.
Oh
I
saw
on
your
Doc's
need
at
least
one
dot.
Oh
cool
cool,
all
right,
so
we've
got
contour
up.
Let's
do
a
couple
more
deployments
and
I'll
see
if
there's
more
questions
to
answer
in
chat
here.
A
Okay,
so
contours
up
and
I've
pulled
up
there,
the
flagger
documentation
for
for
contour
Canary
deployments
and
let's
see
what
we
got
going
on,
so
they
recommend
deploying
contour,
which
I
just
did
so
stuff
on.
This
is
what
I
was
reading.
Contour
one
dot
or
newer
cool.
You
haven't
tried
one
dot,
two
okay!
Well,
we'll
know.
If
things
blow
up,
we
can
always
we
can
always
roll
back
to
one
dot
o
or
one
dot,
one,
no
biggie,
alright!
So
what's
the
first
thing
we
got
going
on
here,
so
install
flagger
using
customize.
A
So
let's
get
this
deployed
and
we'll
take
a
little
bit
of
a
look
at
what
it
creates
and
kind
of
what
some
of
the
pieces
are,
and
maybe
Stephane
can
even
help
us
understand
how
these
pieces
fit
together
here.
So
I
did
notice
that
this
was
a
github
page.
So
I
went
ahead
and
just
pulled
this
down
before
the
show
just
to
make
sure
I
have
the
manifest
so
per
the
per
the
docs.
A
Once
we
do
a
cube
cuddle
apply
with
dash
K
for
customize
and
then
we're
gonna
go
to
the
contour
directory,
which
I'm
guessing
has
flagger
inside
of
it.
So
let's
apply
dash
K
and
we
will
do
the
let's
see
where
did
I
put
this
in
again.
I
think
it
was
just.
Was
it
customized?
It
was
manifest.
What
do
we
got?
Customized
yeah
manifest
customized,
contour,
okay
and
oops.
A
A
We
will
apply
K
right
and
we'll
do
the
customize
or
sorry
manifests
customize
contour,
which
has
these
inside
of
it
unknown
shorthand
flag
K.
What
version
of
cube
cuddle
am
I
using
cube,
cuddle
version,
I
would
think.
Oh
looks
like
K
was
removed,
interesting,
so
my
cube
cuddle
version.
My
client
version
is
113.
How
did
that
happen?
I've
used
customize
before,
but
I
must
have
messed
something
up.
That's
super
interesting,
okay,
so
let's
get
a
let's
get
a
new
version
of
of
cube
cuddle
here,
real
quick.
A
Okay,
so
bring
up
my
browser
here
in
just
a
second
good
catch.
Everyone
I'm
trying
to
think,
though
cuz
I
was
playing
around
with
customize
the
other
week
and
I
just
can't
figure
out
how
I
got
an
old
version
of
it.
Nonetheless,
let's
see
if
this
works
so
we're
gonna,
pull
it
from
the
arch
repository
in
my
case,
so
I'll
do
a
pac-man
RS,
maybe
just
our
cube
cuddle
and
we
will
sudo
that
okay,
oh
I,
put
it
in
there
manually.
A
Okay,
I
must
have
done
something
really
weird,
then
so
I
wonder
if
it's
in
my
user
local
bin,
as
cube,
cuddle,
nope,
let's
see
where
I've
got
it
here,
it's
always
good.
When
the
first
thing
that's
blowing
up.
Is
your
your
cube?
Cuddle?
Okay?
Let's
do
it
live
here.
Let's
say:
let's
go
ahead
and
delete
cube,
cuddle
and
see,
what's
going
on
so
user
in
cube,
cuddle,
cool,
all
right
and
then
I
will
just
go
ahead
and
cheat
here
and
quickly
go
to
this
directory
grab.
A
This
cube
cuddle
bin
from
the
arch
repository
and
then
we
will
try
to
install
it
so,
okay,
alright!
So
while
that's
installing,
let's
check
out
chat
and
see
what
we
got
going
on
all
right
do
ya.
Abc!
Your
question
was:
how
do
people
manage
cluster
deployments?
Cluster
API
and
such
you
know,
I
think
I.
Think
there's
a
lot
of
variance
in
that
today.
To
be
super
honest
with
you,
you
know:
I've
helped.
A
Customers
manage
clusters
in
a
lot
of
different
ways:
some
use
purely
homegrown
systems,
some
use,
you
know
simply
pipelines
and
things
that
try
to
keep
customers
are
clusters
up-to-date.
You
know
that
being
said,
I
think
with
cluster
API
and
with
the
notion
of
like
a
supervisor
cluster
that
could
theoretically
manage
many
clusters,
it
could
simplify
and
or
unify
cluster
management,
but
even
like
a
lot
of
vendors,
you
know
who
make
products
that
you
theoretically
buy
for
X
number
of
dollars.
A
They
kind
of
monetize
off
this
deficiency,
in
my
opinion,
to
some
degree
right
I'm,
even
at
VMware
we're
trying
to
make
that
Multi
cluster
management
a
lot
easy
easier.
So
hopefully
I
answers
a
little
bit,
but
I've
found
I've
personally
found
a
lot
of
variance
and
that's
a
great
question.
Okay,
let's
see
what
we
got
here,
everyone
so
did
we
fix
the
version,
cube
cuddle
version.
Our
client
version
is
117.
Okay,
so
are
we
gonna
be
able
to
deploy?
You
know
just
a
little.
The
you
know.
A
The
demo
gods
and
the
the
presenting
gods
got
to
make
sure
that
we
don't
have
this
too
easy.
Okay,
we
can't
just
we
can't
just
get
away
with
it
working
the
first
time.
So
if
we
do
cube
cuddle-
and
we
apply
kay-
we
will
do
in
this
case-
manifests
oops.
I
should
be
in
the
right
directory
again.
So
let's
go
back
to
the
TGI
K
folder
episodes
106
okay.
So
let's
try
that
apply
one
more
time
with
the
K
much
better.
Okay,
so
see,
I
was
even
playing
with
it
the
other
day.
A
So
it's
got
some
of
the
CR
DS
and
stuff.
So
I
have
no
idea
what
I
would
have
done
to
blow
that
up.
Nonetheless,
we've
got
flagger,
congratulations
everyone,
so
we're
gonna
be
able
to
actually
try
this
out
alright.
So
what
did
this
create
for
us
exactly?
So
we've
got
a
flagger
pod.
Maybe
we
can
start
by
just
quickly
looking
at
that.
So
if
we
get
the,
you
know
the
pod
for
the
namespace.
What
am
I
in
project
contour?
A
A
So
we've
got
this
pod
name
flagger
inside
of
here
it
looks
like
some
of
the
flags
that
got
passed,
so
this
is
probably
where
the
provider
that
kind
of
plugin
detail
gets
put
in
like
I
was
mentioning
like
I
would
think
this
could
be
ISTE,
Oh
or
glue
or
whatever
it
does
look
like.
It's
got
some
built-in
slack
notification
stuff
as
well.
A
So
that's
kind
of
cool,
liveness
probes,
pretty
standard
stuff
readiness,
probes
limits,
okay,
so
it
seems
like,
for
the
most
part,
I'm
not
seeing
anything
too
crazy
in
here
it
looks
like
as
far
as
the
flagger
applique
goes.
It
is
mostly
concerned
about
what
matches
integrating
with
and
then,
where
can
it
find
the
metric
server
at
so
that
all
looks
pretty
pretty
good
alright?
A
So,
let's
see
if
we
can
bootstrap
some
different
tests
now,
one
thing
actually,
let's
check
something
out
here:
every
one
one
thing
I
think
it
did
make:
let's
see
if
I've
got
cute
cuddle,
get
get
I
think
what's
called
canary
okay,
so
canary,
let's
see
if
we've
got
any
I
guess
we
probably
don't
have
any
yet.
It's
probably
gonna
have
me
put
those
in
a
little
bit
later
canary,
API,
okay,
jump
in
ahead
of
myself.
A
So
let's
get
some
of
their
their
recommended
test
data
in
here,
real
quick,
so
we'll
create
a
namespace
for
test
cube,
cuddle,
create
namespace
test,
looks
good
okay
and
then
it
wants
us
to
install
a
load
balancing
service
to
generate
traffic
during
the
canary
analysis.
Okay,
so
this
will
probably
be
a
load
tester
of
sorts,
so
let's
go
ahead
and
apply
that
as
well.
A
Okay,
so
we
will
do
our
apply
this
time
instead
of
customized
contour,
we'll
do
customized
tester,
which
created
service,
flag
or
load
tester
deployment,
flag
or
load
tester
as
well,
looks
good
to
me.
Okay.
So,
let's
see
if
we
can
find
these
real
quick,
they
are
not
inside
of
project
contour,
so
get
all
namespaces
oh
test
test,
namespace
duh.
We
just
created
that
test
namespace.
So
we've
got
the
flagger
load
tester
inside
of
here
as
well,
which
seems
to
make
sense.
Okay,
alright,
and
then
they
have
a
example
app.
A
It
looks
like
we're
gonna
deploy
and
then
maybe
we
can
get
some
of
these
things
interacting.
So
we're
gonna
deploy
a
pod
info
image,
which
is
probably
just
gonna
print
out
data
about
the
running
pod.
So
let's
try
that
now
as
well
and
then
we
should
be
able
to
start
kind
of
trying
some
of
these
bits
out.
So
we
will
customize
apply
again
and
we
will
do
manifests.
Customized
pod
info
looks
good
and
by
the
way,
every
manifest
I
applied
today.
A
I
will
put
in
the
repo
for
TGI
K
episode
106,
so
for
any
reason
you
you're
trying
this
out
while
watching
this
video
be
sure
to
check
out
the
github
page.
These
manifest
will
be
in
here
in
their
current
snapshot
at
date.
So
if
you're
watching
this
in
the
future,
you
you
may
want
to
check
for
newer
manifest.
Nonetheless,
okay,
so
we've
got
a
pod
info
created.
That
seems
fine
with
me.
So,
let's
see
where
pod
info
got
put,
what
do
you
say
so
we'll
check
out
tests?
A
It
looks
like
okay,
pod
info
is
running
in
tests,
so
let's
bring
the
let's
bring
the
text
size
just
a
bit
smaller
here.
Okay,
so
we'll
do
cool
cube,
cuddle,
get
pods
namespace
test.
Alright!
So
here
are
our
pod
info
pods
right
now,
so
just
to
make
sure
we
know
what's
going
on,
let's
also
check
out
the
services,
so
we've
got
a
load
test
or
service
right
now
and
then
maybe
we
can
also
see
if
there's
any
ingress
or
HTTP
proxy
objects
in
here.
A
So,
okay,
good,
doesn't
look
like
anything
yet
so
looking
at
the
service
pod
in
grass.
So
far,
it
seems
like
all
we've
deployed
are
the
pod
info
deployments,
which
is
probably
gonna,
be
our
test
app
and
then
the
flagger
load
balancer,
which
I'm
guessing
will
help
us
generate
load
on
these
as
well.
So
those
seem
to
look
okay
to
me,
I'm
guessing
it
looks
alright
to
you
all
as
well
alright,
so
this
is
probably
the
most
important
one.
So
what
we're
gonna
do
now
is
we're
gonna
bring
in
the
sample
Canary
object.
A
They
have
in
here
and
we'll
take
a
bit
of
a
closer
look
at
this
API
and
maybe
Stephane.
You
can
help
out
if
I
say
anything
terribly
wrong,
feel
free
to
correct
me.
But
let's
take
a
look
at
probably
the
crux
of
this:
all
which
is
the
Canary
API,
so
canary
Gamal,
alright,
just
make
sure
my
paste
won't
mess
me
up.
A
Okay
cool,
so
we
have
got
a
kind
of
canary,
so
this
is
apparently
the
API
that
flagger
is
going
to
introduce
for
us,
given
that
canary
is
attaching
the
pot
info
I'm
guessing
we
create
one
of
these
canary
objects
for
every
type
of
application
which
wouldn't
make
sense.
Okay
and
then
we've
got
so.
It's
well
documented
good.
Ok,
so
we've
got
the
deployment
ref
so
clearly
we're
gonna
be
pointing
at
the
pot
info
names.
So
out
of
curiosity
Stephane,
you
might
know
at
the
top
of
your
head.
Is
this?
A
Oh,
this
is:
does
this
allow
like
label
based
selection?
Or
is
it
strictly?
You
know
going
off
of
names
for
right
now
like?
Could
you
use
arbitrary
labels
here
for
your
deployment
reference
trying
to
think
how
much
complexity
that
would
add,
but
nonetheless
autoscaler
ref,
okay,
so
a
horizontal
pot
autoscaler
to
scale
the
instances
of
pot
info
up
and
down?
So
we
actually
the
t
GI
case
somewhat
recently
on
HPA
same
wrath:
okay,
cool
god.
A
It
makes
sense
to
fajn
same
ref
is
HPA
so
service,
okay,
so
the
service
port
of
where
we're
talking,
I'm
guessing
and
then
the
container
port,
which
also
makes
sense
cool
cool.
It's
got
a
contour
retry
policy
inside
of
here
as
well.
So
I'm
wondering
if
that's
got
something
to
do
with
the
test.
Maybe
we'll
come
back
to
that
and
let's
look
at
the
canary
analysis
first,
so
analysis,
timing,
kpi's,
okay,
so
this
is
probably
where
some
of
the
gold
is
right.
A
We've
got
an
interval
that
is
probably
a
timer
on
how
often
we
do
something
based
on
the
canaries
configuration
I'm
guessing
max.
Okay.
This
is
great,
so
max
number
of
failed
metric
checks
before
roll
back
and
then
the
max
traffic
percentage
routed
to
canary.
So
it
looks
like
our
max
in
this
case
will
be-
will
go
up
to
50%
at
the
very
most
shifting
over
to
a
new
version
of
pot
info,
while
the
tests
are
going
on
and
then
the
step
wait.
Okay,
this
is
cool.
A
So
if
I'm
reading
this
correct
and
again
keep
me
honest,
here's
to
fun
if
I
say
anything
silly,
but
we're
gonna
go
up
by
five
percentages.
I'm
guessing
that's
gonna,
be
on
a
30.
Second
interval
keep
me
honest
there
and
then
our
max
that
will
max
out
on
when
doing
the
canary
analysis
will
be
50%
and
then
I'm
guessing.
If
all
goes
well
at
some
juncture,
we
should
completely
cut
over
if
it
would
automate
it
like
that.
So
inside
of
here,
we've
got
some
metrics.
A
We've
got
a
request:
success
rate
threshold,
request,
duration,
okay,
so
probably
milliseconds,
500,
milliseconds
intervals,
30
okay,
looks
good
and
then
some
Oh
interesting
some
webhooks
okay.
So
this
is
probably
okay
great.
This
is
where
this
this
tester
is
gonna
come
in.
So
let's
see
here,
webhooks
acceptance
tester.
So
this
is
going
to
kick
off
I'm
guessing
immediately.
A
I
can't
tell
if
this
is
gonna,
be
something
that
happens
last
Odo's
is
pre.
Rollout
should
have
read
it
right
there,
so
pre
rollout,
we're
gonna,
do
the
acceptance
test
that
makes
sense,
and
then
we
have
a
web
hook
that
we
can
put
for
load
test.
So
this
is
gonna
kick
off
a
load
test
to
our
application.
In
fact,
I
did
read
ahead
in
the
docs
earlier
today
and
I.
A
Think
I
need
to
put
one
of
my
own
custom
domains
in
here,
just
to
make
sure
that
we
can
make
it
work
and
I
don't
have
to
mess
with
I
won't
have
to
mess
with
my
Etsy
host
file
all
day,
so
I
did
actually
set
up
a
DNS
record
for
today.
Let's
just
make
sure
it's
working,
I
probably
should
have
double-checked,
so
it
was
app
dot,
J,
Kate's
calm,
and
this
is
pointing
to
my
host,
which
is
where
the
Envoy
proxies
are
running.
So
this
should
be
good.
A
Even
it's
a
public
record,
that's
pointing
to
a
private
IP
address,
so,
hopefully
that'll
work.
Let's
go
ahead
and
replace
that
here,
so
cool
get
back
here!
Oops
I
did
not
mean
to
do
that.
So
do
I
solve
my
canary
oops.
Okay
canary
Gamal
hit
the
wrong
key
there.
Everyone!
Let
me
just
bring
over
this
file,
one
more
time:
cool
cool,
all
right!
A
Actually,
I
didn't
even
look
at
what
what
I
did
I,
probably
just
discontinued
the
the
team
accession,
so
I
probably
could
have
reattached.
So
let
me
call
this
CDM,
while
here,
okay
and
then
edit
CDM,
oh
there,
we
go
okay
cool.
So
what
I
was
trying
to
do
here
before
I
hit,
whatever
key
I
hit,
is
to
set
this
to
app
dot
Jake,
it's
calm,
which
will
be
my
particular
IP
address.
So
I'm
curious,
I,
don't
know
Stephane.
If
you
know
I
mean
I'm
sure
you
probably
don't
have
to
top
your
head.
A
A
Apply
canary
dot,
yeah
Mille
oops,
not
that
one
Josh
come
on
now.
My
flags
are
all
messed
up
so
delete
there.
We
go
OK
cube,
cuddle
apply,
this
will
be
CEM,
oh
ok,
think
the
bane
of
my
existence.
This
episode
is
going
to
be
that
a
F
and
K
flag.
Ok,
now
we
should
probably
be
good.
So
if
we
do
a
cube
cuddle
get
canary
for
the
test,
the
test,
namespace
cool,
so
we've
got
one
says:
initializing
name
is
pot
info.
We've
got
probably
a
current
wait
and
then
last
transition
time.
A
I'm
guessing
is
the
last
time
it
tried
to
move
between
two
different
versions,
so
hopefully
that'll,
hopefully
that'll
be
ok.
Alright!
Also,
we've
got
a
canary
we've
got
the
we've
got
the
load
testing
tool
that
will
probably
use
soon
in
there
as
well.
We've
also
got
the
deployment
of
pod
tests
and
contour
deployed
as
well,
so
I
think
we
should
be
looking
pretty
good
now.
This
will
be
our
first
exposure
to
contours
HTTP
proxy.
A
So
before
we
try
doing
anything
crazy
with
an
update,
I'm
gonna
take
their
sample
HTTP
proxy,
and
we
will
create
that
in
the
cluster
and,
let's
just
start
off
by
making
sure
that
we
can
actually
route
traffic,
because
that
seems
like
a
pretty
pretty
important
detail
here.
So
I'm
gonna
copy
that
over
okay
and
we'll
call
this
HTTP
proxy
Gamal,
okay
and
then
we'll
set
pasted
all
right.
A
So
we've
got
pod
info
in
our
virtual
hosts,
as
we
talked
about,
is
going
to
be
app
dot,
J,
Cates,
dot-com,
okay,
so
that
looks
pretty
good
and
pot
info
tests
that
all
seems
to
check
out
pretty
well
all
right
cool
got
that,
so
we
will
apply
the
HTTP
proxy
ml
file
all
right
good
deal.
So
now,
if
we
do
a
cube,
cuddle
get
HTTP
proxy
inside
of
the
namespace
test.
We
now
see
that
there
is
oh.
A
There
was
a
one
from
before
I
thought:
I
checked
for
that,
no
nonetheless,
okay,
this
will
be
our
fully
qualified
domain
one.
So
again,
those
of
you
somewhat
new
to
contour
just
know
that
this
CR
D
is
basically
our
own
ingress
API.
If
you
will,
that
provides
extension
capabilities,
but
if
you're
using
contour
you
have
the
choice,
do
you
need
or
like
those
extra
features?
If
so,
you
can
use
this
HTTP
proxy
resource.
Otherwise
you
can
just
drop
an
ingress
in
here
and
it
would
work,
fine
and
actually
flagger.
A
I
guess
is
a
great
example
of
this,
because
I'm
guessing
not
that
we've
gotten
that
far,
that
flagger
is
going
to
adjust
percentages
of
weight
over
time,
or
at
least
something
like
that,
and
by
using
this
HTTP
proxy
CRD.
It's
actually
gonna
have
those
knobs
available
to
it
more
so
than
it
would
have
inside
of
a
standard.
A
standard
ingress
api
so
seems
to
be
good.
Okay,
let's
just
before
we
try
our
first
canary
rollout.
A
Let's,
let's
see
if
we
can
even
access
this
page,
real,
quick,
cuz
I
have
a
feeling
that
will
be
our
first
point
of
contention.
So
we
know
that
we've
got
this
ingress
resource
here.
So,
let's
start
off
with
just
kind
of
reverse
engineering.
How
some
of
this
would
work?
Okay,
which
is
usually
where
I
start
when
I
do
this
kind
of
stuff,
just
to
make
sure
I've
got
all
my
bases
covered.
So
first
things.
A
First,
if
we
were
to
curl
okay,
first
things:
first,
if
we
end
this
lookup
right,
apt,
Jake
8zo,
do
we
get
oops,
tada
yo
comm?
Do
we
get
the
right
IP
address
which
it
seems
to
be
okay,
so
this
is
the
server
IP
address
that
my
one
of
my
workers
is
running
on
I,
just
know
that,
because
this
is
running
on
a
server
in
the
other
room
and
that's
all
fine
and
dandy,
so
that's
good!
A
Now,
if
we
curl
F
dot,
Kate's,
sorry,
F,
J,
Kate's,
calm,
and
then
we
do
a
V
for
I
guess
we
are
getting
sent
to
pot
info.
So
that's
also
good
news.
I'll
make
that
a
little
bit
smaller.
So
it's
easy
to
read,
but
you
can
see
here
that
the
server
that
it's
receiving
it
as
envoy,
so
that's
a
good
sanity
check.
Even
if
we
weren't
getting
something
back.
Perhaps
we
could
see.
Are
we
hitting
envoy?
And
now
we
seem
to
be
getting
the
pod
info
container,
which
is
great
so
notice.
A
The
version
is
three
dot:
one
dot
zero
at
least
right
now,
so
that
all
looks
good.
We've
got
the
the
cuddle,
clap
logo
which,
since
that
worked
I,
feel
like
we
deserve
getting
to
look
at
that
really
quickly.
So
you
know
that'll
be
our.
This
will
be
our
mascot
for
success
throughout
throughout
today
that
looks
good
and
just
some
details
about
the
Linux
host.
Now
all
right,
so
we're
routing.
That's
that's
a
good
thing
right.
A
So,
let's
see
if
we
can
get
an
update
deployed
and
I'll
just
check
the
the
messaging
real,
quick
waleed,
you
saw
the
bracket
yeah
thanks,
but
I
think
I
got
that
one
sorted
can
use
any
combination
of
Prometheus
metrics
to
control
the
way
traffic
shifts
from
A
to
B.
That's
what
I'm
wondering
to
Maddy
Stephane
probably
knows
offhand
but
I'm
thinking
like
some
really
cool
custom
validation
logic
based
on
things
that
you
can.
You
can
scrape
up
through
Prometheus
that
would
be
really
cool.
A
Oh
okay,
I
see
Stephane
so
flagger
added
that
initial
pot
info
resource,
so
that
was
that
big
flagger
deploy
that
I
did,
which
is
would
make
sense
why
it
didn't
show
up
originally
okay.
So,
let's
see
if
we
can
actually
do
a
promotion
here.
Alright,
so
flagger
implements
the
control
loop
that
gradually
shifts
traffic
to
the
canary
while
measuring
the
key
performance
indicators.
Okay,
so
these
are
the
KPIs
that
we'll
be
looking
at
and
we
saw
some
of
those
I
think
in
the
canary
manifest.
A
So
as
we
do,
this
we'll
try
to
see
if
we
can
figure
out
what's
happening
behind
the
scenes
here
right,
so
traffic
comes
in
the
controller
is
waiting
and
you
know
in
Phase
two
actual.
This
is
perfect
for
us
right,
because
we
know
the
increment
is
five
percent,
so
five
percent
gets
shot
to
v2
and
then
it
increases
the
quantity
of
apps.
It
would
seem,
maybe
through
HPA
and
over
time.
A
It
will
eventually
max
out
at
50%,
which
is
gonna
mean
that
v2
should
be
getting
50%
of
the
traffic
I
believe
I'm
guessing
in
step
5.
If
all
goes
well,
this
represents.
You
got
250,
everything
was
good.
We
promoted
you
fully
to
100
and
then
I'm
guessing.
These
are
just
extra
instances
that
get
killed
off.
If
it's
deployed
too
many
up
front.
Perhaps
so,
we've
got
a
really
easy
trigger
that
we
can
do
here,
which
is
the
pod
info
trigger,
but
I
want
to
make
sure
that
we
can
see
what's
going
on
right.
A
So
underneath
this
look,
wouldn't
it
be
nice
if
we
had
that
events
command
right
now
that
we
were
talking
about
earlier
so
using
cube
Caudill
describe,
we
can
get
the
events
for
this
particular
test,
okay,
so
here's!
What
we'll
do
everyone
will
set
up?
A
watch,
let's
see
describes
pretty
loud,
will
either
set
up
a
watch
for
events
or
describe
and
we'll
see.
If
we
can
watch
this
progress
here
and
then
we'll
trigger
this
and
see
what
impact
it
makes.
A
Or
you
know
what
is
what's
gonna
be
its
signal
that
things
are
going
well
is
what
I'm
wondering
and
if
we
go
back
to
our
Canary
deployment
from
earlier,
it
seems
like
it's
gonna,
be
contingent
on
on
this
right,
the
where
was
it
at
the
webhook
right,
so
I'm
thinking,
it's
gonna
call
this
in
the
flagger
load
tester
and
then,
if
it
gets
to
hundreds
I'm
guessing
it's
gonna
progress
on.
So,
let's
see
if
we
can
tell
tell
what's
going
on
there,
so
give
it
a
shot
all
right.
So
first
things.
A
First,
everyone,
let's
get
a
little
watch
in
place
here.
So
first
thing
in
cube,
cuddle
namespace
is
test,
will
do
a
describe
for
canary
pot
info
okay.
So
here's
what
we've
got
message
wise
so
far,
so
we'll
just
fresh
that
I
wonder.
Maybe
some
of
you
you
linux
geniuses,
do
you
know
if
there's
a
way
to
watch
but
to
always
have
it
start
at
the
bottom
of
what
gets
printed
out
from
the
watch?
That
would
be
kind
of
cool
anywho.
A
So,
okay,
so
we've
got
the
halting
advancement,
observe
deployment
generation,
I'm
guessing
it's
just
from
the
initialization
and
then
the
initialization
is
done.
So
this
is
kind
of
where
we're
at
from
the
describe
event
standpoint.
So
now
let's
go
up
here
and
see
oh
nice
ABC,
so
I
can
just
add
the
after
command
to
grep,
and
then
it
will
do
that
with
the
watch.
That
would
make
sense.
Actually,
so,
let's
see,
if
I
can
do
that
real
quick,
so
cube,
cuddle
get
sorry
describe.
Where
was
that
command?
Add
cube
cuddle
describe
okay.
A
So
if
we
run
that
through
grep
events,
a
30,
then
we
watch
on
that
and
then
I
think
when
it
needs
this
complex
watch
needs
single
quotes,
maybe
oh
cool,
very
cool
nice
to
see
Len
something
new
every
day
from
you
all.
Okay.
So
let's
see
if
these
these
things
get
generated,
cool
protip.
So
if
you
don't
have
any,
if
you
don't
have
the
event
streaming
cube
cuddle
command,
yet
you
can
just
do
at
ABC
recommended
it
looks
perfect
alright!
So
here
we
go
so
we
are
gonna.
A
Do
a
cube
cuddle
set
image,
so
here's
the
command!
Everyone
that
we're
gonna
run-
and
you
know
by
the
looks
of
it-
this
is
going
to
basically
pull
us
through
and
set
the
image
up.
Let's
see,
staphon
you
can
do
a
cube
cuddle
logs
to
play.
Flagger,
okay,
cool
yeah,
good
point:
it's
fun!
Let's,
let's
bring
the
logs
up
to
four
four
flagger
that'll,
probably
give
us
even
more
interesting
details:
oh
it'll
print
out
the
same
events
I
see:
okay,
cool
cool.
A
So
let's
see
if
this
goes
through
with
the
describe
and
worst
case,
we'll
print
out
the
logs
from
there
alright.
So
you
ready
everyone
see
if
it
works
here
we
go
so
we
will
set
the
image.
Image
is
updated
on
our
watch.
We
are
updating,
so
we'll
see
what
we
get
here.
Okay,
so
new
revision
detected
and
scaling
up
pod
test,
cool
cool
and
Estefan
said
in
chat.
We
can
pump
these
same
events
into
slack,
which
is
cool.
A
You
know,
give
us
give
us
some
increased
anxiety
and
our
slack
channels
as
our
new
new
app
version
is
rolling
out
over
time.
Alright,
so
I'm
guessing
at
some
point,
you
know,
based
on
the
docs,
we're
gonna
eventually
see
the
canary
wait.
So
remember,
remember
our
canary
deployment
right.
It
seemed
like
every
30
seconds,
or
so
we
should
start
bumping
up
based
on
whether
this
load
test
I
suppose
went
okay.
Okay!
Here
we
go
first
first
step,
so
we've
got
the
canary
wait
set
to
five.
Let's
try
some
stuff
up
here,
real
quick
everyone.
A
A
So
if
I
send
some
requests
in
so
far,
and
this
kind
of
makes
sense,
I'm
still
hitting
3.1
dotto
nope
I
got
a
3.1
dot,
one
right
there,
okay,
so
clearly
it's
bringing
a
little
bit
of
traffic
in
here
over
time,
but
then,
if
I
hit
it
again,
you
can
see
it's
still
over
time
hitting
3.10
3.10
3.1
dotto.
We've
pushed
our
canary
weight
to
10
okay.
So
what's
this
other
one
halt
advancement,
no
values
found
for
contour
metric
request
pot
info
dot
test
is
not
receiving
traffic
I'm
a
success
rate
possible.
A
So
I
wonder
if
that
stephane
maybe
means
that
the
maybe
the
traffic
triggers
not
kicking
off
to
send
traffic.
In
aside
from
my
traffic,
obviously
which
I'm
clearly
sending
in
at
some
rate
here
first
it
runs
the
conformance
test,
then
increased
traffic,
then
metrics
checks
every
30
seconds,
cool,
okay,
so
far,
I'm
liking.
What
I'm
seeing
I
don't
know
exactly
how
much
traffic
it's
sending
in
via
that
load
test
utility,
but
it
does
seem
to
be
slowly
increasing
me
here.
So
that's
that's
good
news!
A
Oh
it's
Prometheus,
scraping
latency,
okay,
I,
see
I,
see
so
just
a
normal
and
I
guess
that
kind
of
makes
sense.
Since
it
was
just
letting
me
know,
it's
a
warning
there,
okay
cool!
So
it's
sending
that
in
and
all
is
going
well
I'm
gonna,
while
we're
waiting
for
us
to
go
through
I,
think
I
saw
something
in
the
docs
that
shows
how
to
force
a
failure.
A
Let's
see
if
I
can
find
that
again,
because
that
would
be
a
good
thing
for
us
to
validate
after
this
hope,
there's
a
picture
of
what
it
could
look
like
from
the
slack
standpoint.
Okay,
so
we've
got
that
that's
what
I
was
looking
for?
Okay,
so
exactly
to
the
load
test
or
pod,
and
then
we
can
generate
500
errors
on
the
app
generate
some
latency,
okay
cool
cool.
A
So
we
should
be
able
to
hear
everyone,
so
it
seems
like
it
seems
like
if
I'm
reading
this
correctly
we're
sending
the
traffic
in
its
slowly
waiting
us.
It
seems
like,
as
it's
increasing,
that
percentage
it's
going
through
and
actually
you
know
if
succeeding
at
the
requests.
Another
thing
we
could
check
I'm
thinking
it
would
be
reflected
here
if
we
get
the
HTTP
proxy
for
the
namespace
test.
Let's,
let's
see
what
this
has
going
on
with
it
real
quick
test:
oh
yeah,
mol.
A
Okay,
so
we've
got
this
here
and
HTTP
proxy
pod
condition.
Okay,
so
cool
cool
doesn't
seem
to
be
impacting
anything
here.
So
I
wonder:
if
does
it
go
into
the
HTTP
proxy
at
all
and
change
percentages
there
or
where
is
it
changing
the
percentage
of
traffic
shifting
am
I.
Looking
at
that
wrong?
Let's
see
here
so
we've
got
this
pot
info
ingress
I'd
expect
there
to
be
percentage
changes
it's
clearly
doing
it
I'm,
just
not
looking
at
the
right
way,
oh
right
stuff
on
the
other
pot
info
right
right.
A
Okay,
let's
see
if
I
can
get
this
before
it
gets
too
much
farther.
So
this
would
be
the
pot
info
that
flagger
owns
here
right.
So
if
we
do
an
oh
yeah
mall
there
we
go
okay.
So
what
flaggers
got
going
on
here
is
it's
adjusting
the
weight
as
time
goes
on
between
these
two
versions,
so
you
can
see
at
least
at
this
snapshot
in
time.
It's
going
into
this
API
that
contour
respects
and
it's
moved
us
up
to
a
weight
of
40
and
it's
currently
got
the
other
one
at
60.
A
So,
as
we
all
know,
we
should
be
pretty
close
here,
because
once
we
get
50/50
I
think
we
should
be
in
a
place
where
it's
actually
going
to
shift
completely
over
to
the
new
version.
So
let's
go
back
to
our
logs
here.
I'll
just
exit
out
of
this
buffer
advanced
weight
canary
is
25.
Let
me
make
sure
I've
got
the
the
right
version
of
this
advancing
pods
45
good
good.
A
Let's
see
you
said,
ABC
the
DNS
hostname
looked
wrong.
Let's
look
so
cube.
Cuddled
get
HTTP
proxy
for
the
namespace
test,
so
that's
pot
info.
Okay.
Let's
get
that
again
so
pod
info,
oh
I,
see
what
you
mean:
yeah
pot
info,
Canary,
dot,
dot
service,
hi
I,
wonder
if
that
might
be
one.
If
that's
a
bug,
Stephane
says
that
one
is
not
used
and
will
be
when
linker
D
will
support.
Okay,
cool
cool
Hey.
Look
we're
helping
we're
helping
out
you're,
helping
us
out
stephane
we're
helping
you
out.
A
So
we
found
we
found
two
dots
good
job,
good
job,
ABC,
good
catch,
so
yeah.
So
it
looks
like
there's
a
weird
value
in
there,
but
if
I'm
reading
Stefan's
message
correctly,
it
probably
won't
impact
us,
at
least
for
our
use
case
with
just
contour,
okay,
cool
cool.
So
what
else?
What
else
do
we
got
going
on
here?
Right?
We've
got
the
just
get
the
info
here
from
cube:
cuddle
get
out
of
the
watch
mode.
I'm
25
templates
back
to
pod
info
primary
test.
Advanced
Canary,
wait
225.
It
seemed
like
the
Canary
way.
A
It
was
more
than
25,
though
right.
It
seemed
like
it
was
at
40,
so
if
we
do
get
Canary
for
the
namespace,
what
are
we
in
again
test
finalizing?
Okay,
so
clearly
were
we're
closer
I,
wonder
I,
wonder
if
I'm
missing
anything
in
the
events,
the
weight
going
to
25
seems
like
we're
farther
than
that
seems
like.
We
are
nonetheless
and
I
mean
I.
Guess
here's
how
we
could
check
right
if
we
just
get
the
yeah?
Let's
do
it.
Let's
do
a
cube
cuddle
describe
because
that's
what
that's?
A
A
Okay.
So
what
are
we
at
here?
I
think
we
may
have
shifted
over
actually.
So
what
do
we
got?
We've
got
the
the
spec
in
here
host
name
host
name
host
name.
What
are
these
date
target
port?
Oh
I'm,
scrolling.
The
wrong
way,
hang
tight!
That's
why
that
looks
so
that's
why
that
looks
so
messed
up.
That's
not!
That's!
User
error.
Okay,
so
we've
got
the
HTTP
proxy
owner
references
kind,
Canary,
okay,
so
we
can
see
the
owner
ref
from
flagger
here
on
the
HTTP
resource.
A
That's
what
Stephane
was
saying
earlier
that
again
shows
that
this
is
owned
more
or
less
by
flagger,
so
flaggers
won
with
its
hands
on
the
knobs
there,
and
then
we've
got
pod
info
primary.
That
seems
pretty
good
right
and
yeah.
There's
our
validation,
so
weight
is
100,
so
in
theory
we
should
be
completely
moved
over.
So
maybe
maybe
we
deserve
another
another
screenshot
of
our
our
clapping
clapping
cuddle,
but
before
we
do,
let's
just
quickly
make
sure
that
we
have
app
dot.
A
A
So
we
we
saw
some
web
hooks
that
we
could
put
in,
which
seems
like
we
could
put
in
some
arbitrary
web
hooks
based
on
validation
logic,
so
hopefully
you're
starting
to
kind
of
conceptualize
how
this
could
be
pretty
powerful
right,
like
those
web
hooks
our
pluggable,
whatever
test
load,
generating
thing
that
you
want
to
do,
and
then
we
have
the
validating
logic
going
over
time,
stepping
stepping
stepping
and
so
on
now
one
thing
I
would
love
to
do.
I
should
probably
a
quick
time
check.
Okay,
one
thing
I
would
love
to
do.
A
Is
I'd
love
to
see
if
we
can
make
it
fail.
That
would
be
really
cool
to
kind
of
see
the
rollback.
What
do
you
say?
Let's
give
it
a
shot?
Okay,
so
let's
see
what
we
got
here
and
it
looks
like
per
their
docks
by
the
way
a
flagger
team
I
feel
like
people.
Don't
complement
this
enough,
but
good
job
with
your
documentation.
It's
it's
pretty
straightforward,
I!
Think
there's
a
couple.
A
Things
like
I
got
a
little
confused
because
some
of
the
apply
commands
we're
pointing
out
like
where
was
it
at
oh
geez,
now
I'm,
getting
we're
pointing
out
like
a
github
which
I
think
kind
of
assumes.
You
have
the
repo
cloned
and
something.
But
aside
from
that,
these
Doc's
have
been
really
really
great
good
work,
anywho,
so
HTTP
500
errors.
Okay,
so
we're
gonna
exact
into
load
tester,
it
looks
like
and
we're
gonna
run,
a
command
which
is
gonna,
do
status.
A
A
Cool
so
we're
in,
and
if
we
run
this
command
it
should
generate
five
hundreds
now
the
question
is:
does
it
want
me
to
run
another
update?
I
would
think
so.
Oh
yeah,
it's
got
another
update
in
here
too.
Okay
I
see
it
so
we're
gonna
set
this
up.
So
first
thing
we'll
do
is
trigger
this
deployment
all
right.
So
let's
get.
Let's
get
some
windows
set
up
here.
A
So
just
like
last
time,
we
can
kind
of
check
this
out
alright,
so
this
will
be
the
command
we
run
so
we're
gonna
bump
ourselves
up
from
3
1
1
2
3
1
2,
which
makes
sense
we
are
inside
of
the
load
test
or
utility,
so
I'm
guessing.
If
I'm
reading
their
docks
right,
we're
gonna
trigger
the
update.
We're
then
gonna
run
this
command
inside
of
load
tester.
So
that's
gonna
be
this
command,
which
is,
if
I
correct
this.
It
looks
like
this
should
app
JK
its
calm.
A
This
should
start
causing
five
hundreds
to
happen
and
I'm
not
super
sure
how
all
that
works,
because
I'm
not
super
sure
how
load
tester
works,
but
I'm
gonna
take
their
word
for
it
that
this
should
generate
some
failures,
because
that's
the
most
important
thing
is
that
our
canary
gets
failures
back
and
then
it
also
has
a
recommendation
to
generate
some
latency
as
well.
Okay,
so
let's
just
check
this
out.
A
So
when
the
number
of
failed
checks
reaches
the
canary
analysis,
stress
hold,
the
traffic
is
routed
back
to
the
primary
that
contain
the
canary
is
scaled
to
zero
and
the
rollback
is
marked
as
failed.
So
this
is
what
we
want
to
see
everybody
if
we
check
out
the
logs
for
flagger
okay-
and
this
is
what
you
were
saying
earlier-
Stefan's.
So
if
we
tail
these
out
and
we'll
tail
them
out
instead
of
looking
at
the
events,
we
should
see
it
increase
it
increase
it.
A
Increased
I'm,
gonna,
flip
the
switch
to
start
causing
errors,
and
then,
if
all
goes
well
and
I'm
reading
this
right,
we
should
start
seeing
some
type
of
rollback.
So,
let's
see
if
we
can
get
this
to
work
okay,
so
this
will
okay,
so
first
things.
First,
let's
do
this!
Let's
send
this
in
all
right,
so
we've
just
done
an
update
everyone.
So
let's
check
this
out
here
now:
okay,
this
will
be
the
the
key.
We
will
start
watching
logs,
okay
and
hopefully
it
will
say
new
revision
there.
It
is
a
new
revision
detected.
A
Okay,
everyone,
so
we've
got
a
new
revision.
That's
good
new
revision,
now
I'm,
not
sure
if
the
Doc's
just
want
me
to
run
this
right
away
or
if
I
can
run
it
after
a
couple
seconds,
maybe
Stefan
you
can
jump
in
if
I'm
doing
this
wrong,
but
I'm
hoping
I
want
to
see
it
scale
up
like
de
canary
wait.
Five,
maybe
canary
rate
wait,
wait
ten
and
then
I'm
hoping
I
could
run
this
command,
so
it
kind
of
gets
parsh
into
the
rollout.
A
If
I'm
thinking
this
through
correctly
and
then
it
should
start
failing
on
us.
Okay,
so
we've
bumped
up
to
five,
that's
good,
so
we're
basically
doing
the
same
thing
with
flagger.
Now
we
are
headed
up
to
wait.
Five:
hey
Stefan,
my
my
domain
is
accessible
inside
the
cluster.
It's
using
a
public
DNS
record,
which
is
then
returning
a
private
IP
address.
So
if,
as
long
as
that'll
work
I'm
pretty
sure
that
should
be
okay
cool
cool,
so
that
seems
good.
This
is
good
all
right.
Also,
let's,
okay,
it's
got
2:10.
A
Don't
know
if
that
stall
there
is
a
good
thing
or
not.
Maybe
it's
got
something
to
do
with
my
domain,
causing
issues
like
I
said
this
is
returning
a
private
IP
address,
but
the
private
IP
address
it
returns
should
be
accessible
to
whatever
this
is
I.
Don't
know.
If
this
normally
gives
output
the
command
will
generate
504.
Ok,
it
works
check
it
out
everyone.
So
we
are
currently
seeing
some
success
rate
errors.
That's
good!
That's
good!
A
So,
let's,
let's
try
to
calculate
this,
so
we're
at
10
right
now,
success
rate
is
less
is
success.
Rate
is
48.3%.
Is
this
saying
48
3
percent
for
the
P
99
I'm?
Guessing
that's
the
case,
so
our
success
dropped
below
48
point
three
one!
That's
good
and
two
Stefan's
point:
okay,
so
we're
still
getting
errors?
Ok,
good!
So
now
we
stopped.
So
this
will
be
interesting.
Everyone
right
now
the
500s
have
stopped.
The
question
is
gonna,
be?
Will
it
start
picking
up
here
and
kind
of
realizing?
Ok,
things
seem
healthy
again.
A
Let
me
go
ahead
and
bump
up
again
and
if
it
does
then
we'll
just
go
ahead
and
put
a
bunch
of
error
messages
in
more
500s
and
increase
the
latency
and
then
hopefully
we
can
actually
cause
it
to
roll
back.
If
all
goes
bad,
but
I
kinda
want
to
see
it
go
through
both
modes.
So
ok,
so
we're
still
in
it
seems
like
we're
still
in
a
hault
mode.
A
We
haven't
rolled
back
because
we
haven't
triggered
that
condition
yet
based
on
the
amount
of
fail,
but
we
can
see
our
percentages
slowly
going
down
and
I'm
guessing
by
the
next
time
it
triggers
we
might
be
below
we
might
be
below
where
it
needs
to
be
to
continue.
So,
let's
see,
if
that
happens
here,
all
right
yeah.
This
is
pretty
cool.
What
do
you
all
think
chat?
It
looks,
it
looks.
Pretty
good,
looks
pretty
good
success
rate.
Oh
that's
the
success
rate
I'm
looking
at
okay.
That
makes
more
sense,
so
my
success
rate
is
climbing.
A
A
And
then
I'm
thinking
in
another
thirty
seconds,
we
should
kind
of
pick
up
here
and
and
be
good.
Let's
see,
if
that,
if
that
works
out,
okay,
I
guess
I
should
have
lowered
my
time
below
30
seconds.
Okay,
there
we
go
so
we're
healthy
again,
so
we
had
this.
That's
a
perfect
example
of
these
canary
rollouts
right,
we
kind
of
went
forward
and
then
we
started
hitting
500s.
We
got
below
our
threshold,
which
was
set
in
the
canary
object
right
and
it
looks
like
we
halted
halted
halted.
A
A
So
once
the
number
of
failed
checks
reaches
the
canary
analysis
threshold.
Supposedly
we
should
be
good.
So,
let's
just
keep
an
eye
on
this
as
we
go
and
since
we're
getting
a
little
bit
closer
to
time
here,
I
just
want
to
make
sure
we
can
knock
out
one
more
quick
thing
before
we
call
it
so
we'll
keep
this
in
the
background
and
take
a
look
in
a
moment.
So
alright,
so
rolling
out
seems
to
be
seems
to
be
working
pretty
well.
A
The
last
thing
I
wanted
to
check
with
you
all
was
a
little
bit
around
a
be
testing
inside
of
here,
because
I
know
this
is
a
really
common
one,
where
we're
not
just
releasing
kind
of
a
new
version,
but
actually
going
through
and
like
based
on
headers
and
stuff
determining
which
a
version
to
show
to
the
to
the
user.
That's
calling
so
it
looks
like
inside
of
our
canary
analysis.
All
we'll
have
to
do
is
change.
A
Conditions
the
match,
particularly
so
we
should
be
able
to
go
in
here
and
change
some
of
the
headers
keep
that
going
Oh.
What
did
it?
What's
it
upset
about
HTTP?
Oh
I.
Did
the
app
example
again:
Josh
come
on
app
dot,
J
Cates,
we'll
see
if
this
is
a
gonna.
Hang
us
up.
Yep
you're
right
I
had
the
wrong
domain
there.
A
All
right.
Let's
see,
if
that,
if
that
goes
through
all
right,
maybe
we'll
make
it
fail
and
then
ideally
I
think
so.
There's
two
things
I
want
to
see
with
you
all
I
want
to
see
if
we
can
get
an
a/b
test
done.
Real,
quick
and
I
also
want
to
see
if
we
could
see
this
roll
back,
because
what
we
didn't
get
to
hear
just
yet
cuz
I
was
too
I
was
too
excited
about,
seeing
it
halt
and
then
go
again
is
once
it
fails
for
the
amount
of
duration
we've
set.
A
We
should
eventually
get
something
like
this.
That's
like
hey
bad
things
are
happening.
They've
been
happening
for
too
long.
I'm
gonna
start
rolling
back
right.
So,
let's
see
if
we
can
get
that
get
that
triggered
inside
of
here
cool
cool
and
then
we'll
get
to
our
a
B
testing
as
well
yeah.
This
is
very
cool.
A
Okay,
all
right!
There
we
go,
so
we
got
it
canary,
failed
scaling
down
pod
info
dot
test
so
rolling
back
checked.
This
is
our
failed
threshold,
so
we'll
check
our
canary
one
more
time
to
validate
this,
but
it
looks
like
based
on
the
five
that
we've
now
hit.
We
are
going
backwards
and
it
should
bring
us
back
to
I
believe
our
original
amount.
So,
while
that's
going,
let's
check
it
out.
So
if
we
do
a
cube,
cuddle,
cube
cuddle
get
canary
inside
of
the
right
namespace.
Of
course
perfect.
A
Okay,
so
we
see
I
mean
not
perfect.
If
this
was
a
real
thing
but
perfect
in
our
case,
pod
info
has
failed,
so
we
should
be
back
on
purely
version
one.
So
if
we
check
that
out
from
a
contour
perspective
will
do
cube,
cuddle
get
the
HTTP
proxy
for
pot
info
in
the
name,
space
of
test
good
and
then
we'll
do
the
Hamill
version
and
then
we'll
do
less
all
right.
So
now
you
can
see
we
have
wait.
A
100
pointed
to
this
particular
pot
info,
which
is
good,
and
then
that
means
that
if
we
were
to
curl
this
app
yep
there
we
are
so
we
rolled
back
to
3,
1
1,
awesome,
very
cool
everyone,
so
we
successfully
shifted
traffic
induced
an
error
based
on
the
load
testing
tool
which
again
I'm
not
super
familiar
with
what
the
load
testing
tool
does.
But
it's
pretty
clear.
We
triggered
a
bunch
of
500
responses
and
then
it
completely
rolled
back
to
its
original
versions.
That
is
pretty
darn
cool,
very
awesome.
A
Does
that
make
sense
chat?
Does
anyone
have
any
questions
on?
That
seems
seems
pretty
straightforward.
I'm
really
really
excited
about
it.
So
let's
do
one
more
thing
before
we
end
it
here.
Let's
see
if
we
can
do
like
an
a/b
test,
just
to
show
a
change
in
a
change
in
some
of
these
canary
settings,
real
quick.
What
do
you
say
so?
A
A
be
testing
for
those
of
you
who
might
not
be
as
familiar
is,
is
in
most
cases
something
where
you've
got
two
versions
and
then
based
on
some
type
of
criteria
you
go
out
and
it
will
send
you
to
the
right
version
based
on
some
header
or
other
criteria
again
I've
mentioned
earlier,
but
a
super
common
use
case
is
like
I
want
to
roll
this
thing
out.
This
UI
changed
to
two
amia
and
not
the
US
and
based
on
results.
A
I
want
to
determine
whether
it's
something
I
want
to
roll
out
in
a
larger
scope
or
whatever
it
might
be
right.
So,
overall,
it
looks
like
we're,
mostly
just
changing
our
canary
analysis.
Let's
see
if
we
can
do
that
now,
so
we
will
back
up
and
we
will
get
back
into
C
gamma,
which
is
where
we've
got
canary
n.
Let's
find
the
analysis
here,
which
is
right
here.
Ok,
all
right,
so,
ok,
so
actually
here's
some
of
the
things
we
were
just
looking
at
right,
so
threshold
5,
is
seems
to
be
what
we
hit.
A
We
went
up.
We
failed
enough
times
once
we
failed
5
times
we
triggered
a
rollback,
there's
our
30-second
interval,
which
I'm
sorry,
everybody
you
had
to
just
stare
at
me
for
each
one
of
those
30-second
blocks
and
then
there's
our
max
weight
that
we
were
setting
before
now.
It
seems
like
in
this
case
it
is
going
to
want
to
do
this
based
on
a
header,
so
I'm
gonna
copy
this
out
of
there
Docs
and
see
if
we
can
sort
of
put
it
in
here.
A
A
So
it
looks
like
as
long
as
we've
got
that
header
going
in
for
exact
insider,
we
should
be
able
to
see
a
specific
version,
so
what
I'm
trying
to
figure
out
is
if
we
set
a
new
image.
I'm
curious,
what's
gonna
happen
as
far
as
routing
go.
So,
let's
see
here
so
if
we
look
at
where's
the
update
command,
okay
so
trigger
a
canary
deployment
by
updating
the
container
image.
So
we
go
2
3,
1
3
that
makes
sense
we're
gonna,
bump
ourselves
to
3
1,
3
and
then
I
would
think.
A
We'd
have
3
1
&
3,
1
3
running
side
by
side
now,
flagger
detects
a
deployment,
revision
changed
and
starts
the
a/b
test.
Ok,
so
canary
iteration
iteration
iterations.
So
this
is
the
the
30
second
intervals,
I'm
guessing
well,
I!
Guess
we'll
see
so
waiting
for
rollout
to
finish
so.
This
is
just
like
a
timed
a
be
test
if
I'm
thinking
of
this
correctly,
so
my
hope
would
be
if
we
run
this.
A
A
That's
where
my
that's,
where
my
brain
has
gotten
me
so
far.
So,
let's,
let's
see
if
that's
actually
actually
right.
Okay.
So,
let's
see
what
else
we
got
going
on
in
chat
before
I,
try
that
it
may
reduce
okay,
maybe
testings
specific
client
client
will
always
see
a
others
while
lead.
You
said
an
a/b
testing.
Can
you
specify
by
client,
like
a
specific
client,
always
sees
a
others?
Cb
I
would
think
you'd
be
able
to
do
that.
Walid
I'm,
not
sure
how
advanced
it
is
to
fajn
could
probably
help
us
out
here,
but
I'm
thinking.
A
In
this
case
only
clients
that
satisfy
this
match
are
probably
going
to
get
my
new
version.
I,
don't
know
if
you're
looking
for
a
qualifier
as
well
for
the
old
version
like
having
a
match
for
going
too
old
or
maybe
even
running
an
ABC
test
right,
but
I
would
think
you
could
probably
use
any
arbitrary
header,
like
ABC
said
with
the
user
agent
like
I,
don't
see
why
you
couldn't
use
a
user
agent
and
then
based
on
the
user
agent
route
to
a
certain
pod.
It
seems
like
as
long
as
your
match.
A
Functionality
here
was
set.
It
seems
like
it
would
work,
fine,
honestly
cool.
So,
let's
see
you're
asking
if
you
can
use
flagger
with
tray
traffic
I
have
not
heard
of
traffic
support
but
stephane
in
probably
let
you
know
if
that's
at
all
something
they're
thinking
about
lead
and
may
be
reduced
by
30-second
interval,
okay,
cool
cool
and
then
Stephane
I'm,
not
sure,
I'm
tracking.
Your
suggestion
on
you
forgot
to
change.
The
interval.
Do
I
need
to
change
the
interval
for
canary,
well,
I,
guess,
yeah.
The
interval
30
seconds
would
be
way
too
short
right.
A
Are
you
suggesting
that
I
like
bump
this
up
to
like
five
minutes
or
something?
Is
that
more
or
less
what
you're
saying
probably
would
make
sense?
So,
okay,
so
theoretically,
we'd
be
open
for
five.
Five
minute
interval
with
canary
analysis
seems
good
and
then
Stefan's
head
yeah.
So
my
my
match
logic.
Is
it
as
advanced
as
contours
HTTP
proxy
filters?
Okay,
great
points!
So
let's
see,
let's
see
if
it
if
it
works
here,
so
no
op
changed
the
min/max
two
iterations,
not
sure
I.
Follow
you
there
Stefan
I'm,
guessing
I'm,
doing
something
wrong
non.
A
A
Don't
see
anything
crazy
here,
every
one,
so
it
seems
like
web
hooks
insider.
That's
just
a
cookie
example.
We're
doing
this
one.
So
interval
is
1
minute
thresholds.
5
iterations
is
10,
do
I,
set
it
eration
x'
before,
given
the
other
I,
guess:
I,
wouldn't
right,
yeah,
so
iterations
of
so
I'm
having
a
hard
time
wrapping
my
head
around
all
of
the
API
just
cuz,
it's
new
to
me,
but
I'm,
guessing
iterations
would
be
if
I
put
a
10
here.
I
would
run
this
a
B
for
10
iterations
of
5
minutes.
A
Let
me
know
if
that
logic
is
correct,
so
I'm
thinking
this
means
10
iterations
of
5
minutes
where
the
a/b
test
is
live.
So
now
we
just
got
to
see.
Can
we
match
this
thing
and
contour
as
far
as
contours
api
is
concerned,
so
we'll
put
flag
or
eon
contour
to
the
test
here
as
our
as
our
grand
finale
everyone?
What
do
you
say?
Okay,
so
remember,
our
header
is
X
canary
I'll.
Try
not
to
forget
that
and
then
exact
match
for
insider
good
deal
all
right.
So
first
things.
A
A
We
should
be
good
now.
Is
that
long
for
us
now
for
five
minutes?
That's
a
good
point.
Well
lead.
If
we're
gonna
see
if
we're
gonna
see
the
iterations
happen,
you
might
not
want
to
wait.
Five
minutes
per
iteration,
it
seems
like
I
might
be,
might
be
making
that
a
little
too
long.
I
agree
with
you.
Let's
drop
it
down
to
let's
drive
down
to
30
seconds
I
think
that
seems
seems
pretty
solid
right,
yeah,
pretty
good.
Okay,
Stefan
agrees.
A
Okay,
you
all
don't
just
want
to
sit
here
for
five
minute
intervals
in
between
each
one.
I
mean
that
seems
seems
totally
reasonable
to
me.
Okay,
so
we've
got
that
set
up.
I
think
I
think
we're
good.
Oh
the
load
test
the
load
test
right,
so
yeah
should
I
change
the
load
test.
I
guess
I
should
if
I'm
gonna
qualify
it
some
of
it
because
I
was
thinking
like.
Oh,
we
just
used
curl
and
test
it,
which
is
totally
fine.
We
should
still
do
this,
but
I
probably
changed
the
load
test
to
right.
A
So
inside
of
the
load
test
for
my
application,
which
is
right
here,
I'm
going
to
add
a
header
and
in
their
docks,
they've
actually
got
snippet
of
that
for
me.
So
I'll
steal
that
to
save
us
some
time
here.
What
do
you
say
so
right
before
host,
not
that
it
would
matter
I'm
gonna,
put
in
X
canary
insider,
so
I
will
be
passing
in
this
test
here:
okay,
just
looking
it
over
one
more
time
in
chat.
If
you
see
any
mistakes,
let
me
know
but
I
think
overall,
this
looks
good.
A
So
all
should
be
well.
Okay,
let's
save
it
up
and
let's
see
what
happens
now
so
we'll
apply
this
one
more
time
to
get
it
configured.
Okay,
we'll
do
a
cube
cuddle!
Okay,
actually
we'll
keep
it
as
is
for
now.
So
let's
see
what
we
can
get
run
in
here,
so
first
things.
First,
let's
run
the
logs,
so
I
copy
and
paste
it
out
of
maybe
some
other
guide.
A
I,
don't
know
if
that
that
may
have
gotten
copied
weird
so
namespace
test
where's
my
where's,
my
logs
command
from
before
project
contours,
where
I
had
it
yeah
so
I
think
very,
very
minor
stuff
on,
but
for
your
docks,
I
think
in
the
a
be
testing
section,
I
think
you've
got
an
example
that
probably
got
pulled
from
the
app
mash
that
mesh
thing,
but
that's
no
big
deal.
Okay,
so
we're
we're
logging,
so
HTTP
proxy,
updated
sync
test
info.
A
A
All
right,
so
it's
seen
that's
that's
a
good
thing,
we're
so
close
everybody,
let's
see
if
we
can
make
this
work.
Okay,
so
once
it
starts
its
thing,
I
would
think
now,
it's
probably
in
a
B
mode.
So
rather
than
seeing
percentages,
go
up,
I
think
we
should
just
see
iterations
ping
every
30
seconds
and
I'll
start
messing
with
curl
up
here,
while
we're
waiting
too
so
I
think
if
I
did
I
mess,
this
up
did
I
change.
The
hostname
yeah
I
bet
I
didn't
change,
the
hostname
did
I
know
I.
Did
it's
good?
A
A
Okay,
so
seems
good.
So
now
we'll
curl
the
insider
and
hey
check
it
out.
Everyone
pretty
sweet.
So
if
all
goes
well
now,
hopefully
I
don't
get
excited
over
nothing.
If
we
pull
the
header
for
insider
out,
we
should
now
be
at
three
one
one
right
because
remember
the
old
rollback
failed.
So
if
we
now
go
back,
fingers
crossed
everyone,
we
got
this
we're
just
gonna,
take
the
header
out
and
we
will.
Let's
just
do
this
for
version
no
URL
specified
what
the
heck
did.
A
I
just
do
the
output
crap
for
version
I'm
doing
something
stupid:
okay,
oh
there's
the
H
flag!
Still
there!
That's
why
my
back
here!
So
let's
do
this
for
version
actually,
last
T
gik
that
I
did
or
a
couple
ones
ago
someone
told
me
I
should
be
using
rest
and
peace
crap,
because
that's
the
new
cool
stuff
right
all
right
there
we
go
check
it
out.
So
we've
got
three
one
one
and
when
we
add
in
the
header
again
and
do
another
RG
and
do
version,
we
get
three
one:
three
pretty
sweet.
A
Everyone
check
it
out,
so
we're
going
through
our
iterations
over
time,
really
really
cool,
that's
sweet
with
that
I
mean!
Obviously
we
deserve
this
now
right.
So
we've
got
our
canary
iterations
going
I'm
guessing
if
B
started.
Getting
some
failures.
Remember
we're
still
we're
still
using
the
webhook,
so
we're
still
triggering
the
load
test.
A
So
if
B
started
getting
failures,
the
a
clients
wouldn't
be
affected
in
the
first
place,
or
you
know
the
clients
without
the
header
and
then
B
would
start
failing
and
I'm
thinking
that
during
these
iterations,
if
we
get
over
our
threshold
of
failures
similar
to
last
time,
we
would
then
roll
back
and
I'm
guessing
roll
back
in
this
case
would
simply
mean
that
we
would
just
go
back
to
the
first
one
exclusively,
regardless
of
your
headers.
So
that
is
super
super
cool,
I
say
the
last
thing
we
check
out
here,
everyone
and
then
we'll
we'll.
A
You
know
just
high
level
everything
and
call
it
call
it
a
call
it
a
day.
Let's
do
a
quick
cube,
cuddle
get
get
HTTP
proxy
for
the
namespace
test.
This
was
pod
info
and
we'll
do
an
oh
yeah
Mille
out
to
less
all
right.
So
what
do
we
got
set
up
here?
Everyone
so
check
it
out.
We
have
got
the
no.
Those
are
the
request
headers
from
before.
Where
is
my
request?
Header
logic
at
pod
info
I'm,
looking
for
insider
in
here,
I
guess
insider
should
be
written,
doesn't
say
anything
about
insider.
A
Am
I
messing
so
there?
It
is
I,
don't
know
why
what
I
did
there
anywho?
So
here's
the
condition
it
added
to
the
spec.
So
again,
this
is:
what's
super
cool
about
you
know,
flagger
being
in
your
workflow.
Is
that
it's
plugging
into
you
know
the
whatever
api?
Is
that
it's
aware
of
and
has
plug-ins
for?
I
don't
know
if
it
uses
a
plug-in
model
or
what
the
deal,
but
what
the
deal
is,
but
nonetheless
it
sits
in
the
cluster
and
it's
like
hey
I
can
talk
to
this.
A
Do
I
can
talk
to
glue
I
can
talk
to
contour.
Just
declaratively.
Tell
me
what
your
rollout
strategy
is
and
I'll
turn
the
knobs,
because
I
understand
the
API
and
I'll
make
it
happen
which,
as
we
all
as
a
group,
get
deeper
into
like
the
get
ops
realm
and
we
start
doing
more
stuff
with,
like
you
know,
our
go
flux
or
whatever
those
those
projects
are
and
in
the
grand
scheme
of
things
I
know.
There's
like
been
talked
about.
A
Some
good
ops
projects
merging
and
whatnot,
but
nonetheless
is
we
get
more
into
the
conceptual
get
ops
model.
I.
Think
things
like
this,
be
it
flagger
or
otherwise,
although
it
looks
like
flaggers
off
to
a
great
start,
are
just
gonna,
be
super
super
critical
in
making
sure
that
app
teams
feel
comfortable
with
making
a
change
having
it
slowly
roll
out
and
how
cool
is
it
that
it's
fully
automated
has
rollback
mechanisms
and
all
that
good
stuff?
A
This
is
some
really
really
neat
stuff,
so
to
kind
of
end
it
here,
I
just
want
to
look
in
chat
and
make
sure
we've
got
any
all
questions
answered,
so
drop
the
H
yep.
You
guys,
you
guys
got
it.
You
all
got
it
Maddie.
How
do
you
automatically
add
that
header
to
a
small
percent?
That's
a
good
question
Maddie,
so
it
Stefan.
A
If
you're
still
on
I,
know,
we've
been
here
for
for
a
long
time
could
could
we
get
as
specific
as
saying
hey
go
ahead
and
do
it
for
the
insider
header,
but
only
on
this
percentage
of
traffic?
On
top
of
that,
in
other
words,
can
we
kind
of
like
embed
canary
analysis
conditions
to
some
degree?
Is
that
something
that's
able
to
be
done
today
or
being
thought
about?
That
would
be
super
super
interesting,
so
very
cool.
A
So
those
of
you
who
dig
a
little
bit
deeper
into
flagger
I,
obviously
didn't
exhaustively
cover
everything
it
can
do,
and
some
of
you
might
be
using
different
different
service
routing.
So
if
using
nginx
app
mash
sto
be
sure
to
check
that
out,
it's
it's
obviously
gonna
be
capable
to
work
with
a
lot
of
your
different
things.
Some
other
things
that
I
just
want
to
mention
as
well
and
call
out
to
the
project
since
I
didn't
explicitly
try
it
today
is
that
it
also
supports.
A
Where
is
it
at
I'm?
Looking
for
the
blue
green,
the
blue
green
mention,
oh
there
it
is
okay,
so
it
also
supports
blue
green
traffic,
switching
which
I
did
not
show
you
all
today.
So
just
be
aware
that
that's
there
and
let's
see,
if
there's
anything
else,
I
wanted
to
mention
I.
Think
that
just
about
does
it.
So
unless
anyone
has
a
question
in
chat,
hang
out
for
another
moment
here
once
again,
we
got
it
everyone
we
made
it
happen.
So
we
tried
out
Canary
deployments
shifting
traffic
over.
A
We
failed
a
canary
deployment
and
got
to
see
it
roll
backwards.
We
did
this
all
with
contour
one
two
by
the
way
which
wasn't
tested
at
all.
So
thank
goodness
that
worked,
and
we
even
saw
an
a/b
test
where
we
were
able
to,
based
on
a
header,
send
traffic
to
a
certain
version
of
the
application
so
again
kind
of
the
call
to
action.
A
A
So,
alright,
everyone
thanks
so
much
chat
thanks
for
hanging
out
and
finding
my
typos
and
answering
questions
Stephane
if
anyone's
still
in
chat,
please
be
sure
to
thank
Stephane
from
the
weave
works
project
who
spent
a
big
chunk
of
his
afternoon
or
evening
here
help
this
out
and
answering
questions
we
we
super
appreciate
it.
Stephane
keep
up
the
awesome
work
with
flagger
and
I
guess:
you'll
be
seeing
Joe,
Duffy
or
myself
next
week.
So
we'll
see
you
all
then
have
an
awesome,
Friday
and
an
awesome
weekend.
Everyone
thanks
again
bye.